HyTorch - A lisp flavoured PyTorch

Clearly, this post falls into the category: "don't try this at home", or, the developers' equivalent: "don't use this in production". With the warning out of the way, you should really try HyTorch if you feel particularly bored on a cold Sunday!

"HyTorch" is not really a thing or library. It is just hy, "a dialect of Lisp that's embedded in Python", and PyTorch 1.0 nightly, because we don't want this project to be too stable.

Why am I doing this? Don't get me wrong, I really like python, and it's my main language for all things data science and machine learning. (Also there are pretty cool things happening in python land: type annotations (PEP484, PEP483, and PEP526), PEP557 data classes, hypothesis - property based testing, asyncio and trio, dask, and more). But I also like functional languages, immutability, "proper" concurrency and parallelism like Clojure, Elixir/Erlang with OTP and more. Sometimes it's just fun to play with a new thing (even though in this case it's technically still python).

Hy lang

We setup hy in a virtualenv for hy:

virtualenv --p python3.7 .venv
source .venv/bin/activate
# install hy
pip install git+https://github.com/hylang/hy.git

Here is a short demo of a few very basic hy features:

#!/usr/bin/env hy

(print "Hello world")
(print (+ "hello" "world" "!"))

(defn greet []
  (print (* "Hello world!\n" 5)))


; lists
(setv some_list [1, 2, 3])
(print some_list)

; dicts
(setv some_dict {1 "one" 2 "two"})
(print some_dict)

; using python's std lib
(import os)
(print (os.cpu_count))

You can also import hy from your python program! Pretty crazy already :)

If you would like to learn more, the official hy tutorial and learn x in y are good. Paul Tagliamonte, one of the creators of hy, has a very entertaining talk on YouTube: Getting Hy on Python: How to implement a Lisp front-end to Python.

PyTorch nightly

Install the PyTorch 1.0 preview with pip for linux and python 3.7 and CUDA 9.0:

pip install numpy torchvision_nightly
pip install torch_nightly -f https://download.pytorch.org/whl/nightly/cu90/torch_nightly.html
python -c "import torch; print(torch.cuda.is_available())"
# True

Hy + PyTorch = HyTorch

Now let's do some transfer learning with HyTorch. We'll write a little image classifier that classifies dogs and cats.

 # start the hy repl
hy 0.15.0+48.gc5abc85 using CPython(default) 3.7.0 on Linux
=> (print "# General infos")
# General infos
=> (import torch)
=> ; this is already amazing!
=> (print "GPU available:" (torch.cuda.is_available))
GPU available: True
=> ; we'll mode the model to the GPU later
=> (setv device (torch.device (lif (torch.cuda.is_available) "cuda:0" "cpu")))
=> (print "Device:" device)
Device: cuda:0
=> (print "# Get the data")
# Get the data
=> ; I copy and pasted some utility functions
=> ; from an old pytorch tutorial into utils.py.
=> ; I'm importing python here!
=> (import utils)
=> (setv [train_dl val_dl]
...   (utils.get_data :batch_size 128 :image_size 224 :sample False :download False))
Loading data from data/dogscats/train.
Loading data from data/dogscats/valid.
=> (print "# Get the model")
# Get the model
=> (setv module (utils.get_model 2))
=> ; this is a simple frozen ResNet model.
=> ; Now we use the amazing PyToune to train the model.
=> (import [pytoune.framework [Model]])
=> (setv model
...   (Model module "adam" "cross_entropy" :metrics ["accuracy"]))
=> (setv model
...   (model.to device))
=> (print "# Train...")
# Train...
=> (import pprint)
=> (pprint.pprint
...   (model.fit_generator train_dl val_dl :epochs 2))
Epoch 1/2 52.61s Step 180/180: loss: 0.149578, acc: 94.730435, val_loss: 0.067275, val_acc: 97.750000
Epoch 2/2 53.10s Step 180/180: loss: 0.083727, acc: 96.821739, val_loss: 0.065457, val_acc: 97.250000
[{'acc': 94.7304347773013,
  'epoch': 1,
  'loss': 0.14957752645534017,
  'val_acc': 97.75,
  'val_loss': 0.06727455681562423},
 {'acc': 96.82173912247367,
  'epoch': 2,
  'loss': 0.08372681827130525,
  'val_acc': 97.25,
  'val_loss': 0.06545671501010657}]

Wow, this is cool and it all worked out of the box. Downloading pytorch took me longer than writing the code to train the model and I haven't used hy in several years.

I don't know where I want to go with this, but it sure was a fun little experiment :)